GTC Berlin 2026 | NVIDIA | Jensen Huang | Building Europe's AI Economy
1. The Real New Product of This Keynote Is a Container
In 2016 NVIDIA delivered the first DGX-1, and the unit of compute was a node. Grace Blackwell NVL72 had the whole rack assembled inside the supply chain, and the unit became the rack. Berlin pushes one layer further: the whole cluster goes into a shipping container.
Jensen Huang's story is consistent: every time the unit of compute changes, the supply chain gets rewritten. NVL72 taught every computer maker in the world to build racks; DSX POD asks the ODMs to ship a small AI factory. NVIDIA sums it up in three lines: factory-built, integrated, pre-tested clusters; delivery-to-deployment cut from months to days; the container is the AI factory.
Put simply: the unit NVIDIA sells just got bigger again, from GPU to server to rack, and now to container.
▲ Vera Rubin DSX POD is a factory-built, pre-tested modular AI factory reference architecture. The bottom shows racks and cabling inside the container; the top lists 17 partners. (Image source: NVIDIA | GTC Berlin 2026 keynote)
2. Why Now: Start a Quarter Late and $10B of Revenue Slips
Huang offered numbers that are easy to work with: a 1 GW AI factory costs about $50B all-in (land, power, shell, compute, networking, storage and cooling) and generates tens of billions of dollars of revenue a year. Assume $40B a year, and a three-month delay pushes back roughly $10B of revenue.
He estimates that going from equipment arrival to the first token takes 4 to 5 months today; moving integration into the supply chain aims to bring that down to about one month. In the meantime, GPUs that have already been paid for sit idle. DSX POD is not solving for performance; it is solving for idle capex.
We saw the same logic in our OCP Global Summit 2026 breakdown of DSX MaxLPS, where NVIDIA claws back wasted megawatts on the power side. In Berlin it claws back wasted months on the time side. Within the same DSX framework, NVIDIA's video also says MaxLPS uses a dynamic control loop to steer power to the workload and delivers the same output with 30% less power.
3. Breaking Down DSX POD: 17 km of Cabling Moves Into the Factory
A Vera Rubin DSX POD consists of 7 liquid-cooled NVL72 compute racks, linked by 17 km of Spectrum-X cabling (on stage Huang first said 17,000 km and immediately corrected himself; the official video says 17 km).
Item | DSX POD spec / NVIDIA's description |
Compute | 7 Vera Rubin NVL72 racks, fully liquid-cooled |
Scale-out cabling | 17 km of Spectrum-X |
Power | 800V DC delivered to the racks |
Cooling | Closed-loop liquid cooling, coolant up to 45°C; dry coolers replace evaporative cooling |
Water | At 2 GW, saves 54 million liters of water a day, about the daily use of 180,000 households |
Deployment | Once site utilities are ready, from months to days |
Partners | Anduril, Armada, Bull, Crusoe, DayOne, Dell, Delta, Eaton, Gigabyte, HPE, Ingrasys, Lenovo, Pegatron, QCT, Schneider Electric, Vertiv, Wistron |
STT estimate: 7 racks × 72 GPUs = 504 GPUs. Spread 17 km across them and you get about 34 m of scale-out cabling per GPU. That is only inside the POD; it excludes fiber between PODs and to the backbone.
For the optical supply chain, what matters more than the quantity is the location:
These 17 km of cabling, termination and testing used to be done mostly on site by cabling contractors; now they are done inside the ODM's factory and burned in before shipping.
Demand for pre-terminated fiber assemblies, high-density connectors and outgoing optical test stations will concentrate inside ODM plants.
Spec and purchasing decisions shift toward the ODM. That matches what we found in OCP Global Summit 1/6 | NPO Wins on Buying Power, Not Volume: whoever controls integration controls the purchase of optical components.
The same shift is happening at the node level. Huang said a Grace Blackwell node took about two hours to assemble, with cables everywhere and about 80% liquid cooling; a Vera Rubin node has no cables, is 100% liquid-cooled and assembles in under 5 minutes. The cables inside the node are gone and the cables between racks are pre-installed, so labor and uncertainty are pulled out at both the node and the data hall.
4. Breaking Down CPO: From Showpiece to Standard in the Switch
Huang brought the Spectrum-X switch on stage and pointed to the lasers and optics: silicon photonics are attached right next to the switch chip using COUPE, the process NVIDIA developed with TSMC, so signals no longer run across a long stretch of PCB or through pluggable transceivers. The spec is 512 ports × 200 Gb/s, i.e. the 102.4T class, and in his words it is in full volume production. He gave no new power, yield or shipment figures, and most of the specs were already disclosed at GTC 2025; what Berlin actually confirmed is volume production.
Which Switch This Is
512 × 200G corresponds to the 2U liquid-cooled SN6810, which can also run as 128 × 800G. The larger SN6800 is 5U with 4 switch chips and 409.6T of total bandwidth, configurable as 512 × 800G or more than 2,000 ports of 200G. NVIDIA originally scheduled the InfiniBand Quantum-X Photonics for 2H25 and the Ethernet Spectrum-X for 2H26; Berlin confirms Spectrum-X has entered volume production on schedule.
Architecture Breakdown
Item | What NVIDIA has disclosed |
Process | TSMC COUPE |
Modulator | Microring modulator (MRM), 200 Gb/s per lane |
Optical engine | 1.6 Tb/s each |
Light source | Centralized external laser source (ELS) feeding all optical engines, about 75% fewer lasers than pluggables |
Power | Initially claimed about 3.5x lower than pluggables; the 2026 claim is up to 5x network power efficiency |
Reliability | MTBF claimed at 10x that of pluggables |
Assembly | Optical engines soldered to the package substrate, compatible with standard production flows; fiber connectors plug in from the top |
STT estimate: 102.4T ÷ 1.6T gives about 64 optical engines per SN6810. Applying Quantum-X's ratio of 36 laser inputs for 18 engines, that would be about 128 laser inputs. This is a ratio-based estimate, not a figure NVIDIA has published for Spectrum-X.
Moving the laser outside is what makes this architecture serviceable: the part most prone to degradation becomes a replaceable module, so a switch does not have to come offline because a laser ages.
Who Does What in the Supply Chain
Publicly named roles: TSMC fabricates the silicon photonics chips, SPIL handles chip-level packaging and test, Lumentum and TFC Communication supply lasers, Foxconn handles system R&D and assembly, and NVIDIA does final testing before shipment; Corning and Senko are also on the partner list.
Three Signals for Optics
CPO lands first in scale-out, in line with our long-held view. On the scale-up side, the NVLink switch tray shown here is still in-rack interconnect for 72 GPUs, and there was nothing new on when optics will enter NVLink.
Optical value shifts from transceivers to optical engines, lasers and fiber. The front panel has no transceiver cages, only fiber connectors and laser modules; add in-factory pre-cabling for DSX POD, and fiber density and connector yield become the integration challenge, growing another 4x at 409.6T. Compare our analysis of how Corning fits 1,152 fibers into 1U.
External lasers are the most direct incremental demand. Lumentum and TFC Communication are the only laser suppliers publicly named so far; when a second source qualifies, and how many channels at what power each switch needs, are the next things to watch.
5. Agents Split the AI Factory Into Four Kinds of Nodes
The other main thread was agents. Huang broke an agent into five elements: the model, the harness (which he calls the digital exoskeleton), the sandbox, knowledge and tools. On the hardware side these map to four kinds of nodes:
Vera Rubin NVL72 compute node: runs the model
Vera MGX compute node: runs the agent harness in an isolated environment
Vera CMX storage node: serves knowledge retrieval
Vera RTX compute node: runs CUDA-X tools (simulation, solvers, data processing)
▲ NVIDIA splits the agent workload into four kinds of nodes: model, secure agent, knowledge and tools. (Image source: NVIDIA | GTC Berlin 2026 keynote)
The demand numbers are blunt: weekly agent token usage on OpenRouter went from 3.9T in early February to 132.6T in early October, about 34x. On performance, SemiAnalysis InferenceX preview data shows Vera Rubin NVL72 producing about 30x the tokens per $1 of TCO of GB300 NVL72 on a DeepSeek V4 Pro agentic workload; Huang also said Blackwell was about 30x faster than Hopper and Vera Rubin is another roughly 30x in two years.
▲ Weekly agent token usage on OpenRouter: 3.9T in early February, 132.6T in early October, about 34x. (Image source: NVIDIA | GTC Berlin 2026 keynote; data from OpenRouter)
What it means for networking: agent workloads are not only GPU to GPU. Harness calls, knowledge retrieval and tool calls bounce among CPUs, storage and DPUs, so east-west traffic now has more kinds of endpoints. That is why NVIDIA put Vera CPU and BlueField-4 front and center: Sentry, the hardware monitoring layer of the OpenShell secure runtime, runs on BlueField-4. We wrote at GTC Taipei 2026 that in the agent era optical connectivity is the critical path to revenue; Berlin filled in the node types along that path.
6. Risks and Counterarguments
"Days" has a precondition: NVIDIA's claim applies once site utilities are ready. Grid connection, substations and cooling plumbing still take months or more; the container shortens only the last stretch.
Containers don't suit every site: 800V DC, full liquid cooling and dry coolers (45°C coolant needs a suitable climate) all depend on location, and large hyperscalers have their own facility standards and may not adopt DSX POD wholesale.
The 30x is preview data: SemiAnalysis itself notes results may change as validation continues, and the workload was chosen by NVIDIA.
Integration power concentrating at the ODM is not necessarily good news for smaller cabling contractors and independent fiber-assembly makers; if ODMs bring cabling in-house, pricing pressure will rise.
7. What It Means for Taiwan's Supply Chain
ODMs go from building racks to building containers: Ingrasys (Foxconn), QCT (Quanta), Wistron, Pegatron and Gigabyte are all on the DSX POD list. Unit value rises another level, but pre-shipment test and burn-in capacity and floor space have to keep up.
Power and cooling: Delta is on the list. 800V DC to the rack, dry coolers and closed-loop liquid cooling are spec upgrades for the power, CDU and heat-exchange supply chain.
Fiber cabling and connectors: with 17 km of cabling per POD pre-installed at the ODM, Taiwanese makers of MPO jumpers, pre-terminated assemblies and high-density connectors will see their customer shift from cabling contractors to ODMs, and their qualification path change with it.
The CPO production chain: Spectrum-X CPO volume is tied to COUPE, with TSMC making the silicon photonics and SPIL handling packaging and test, while back-end coupling and test remain critical links.
Conclusion
On the surface this Berlin keynote was about Europe's AI economy, agents and robots. For optics there was really one signal: the integration point of the AI factory is moving from the data hall into the ODM's factory. NVL72 moved the rack into the supply chain; DSX POD moves the 17 km of cabling between racks in as well; and the same keynote declared CPO switches in full production. Over the next year, where optical components ship, where they are tested and who decides what to buy will all move together.
Three things worth watching next:
When the first DSX PODs actually ship, and to what kind of customer (neocloud or hyperscaler).
Whether ODMs bring fiber assemblies and outgoing optical test in-house, or leave them to existing cabling suppliers.
The share of Spectrum-X CPO switches inside each DSX POD.
This article is for technology and industry trend analysis only and does not constitute investment advice.
Comments